Abordagens neurais para controle de conteúdo pornográfico

• Autor(a) principal: Simões, Gabriel da Silva
• Data de Publicação: 2019
• Tipo de documento: Tese
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da PUC_RS
• Texto Completo: http://tede2.pucrs.br/tede2/handle/tede/9162
• topic: Neural networks; convolutional neural networks; classification; detection; pornographi; censorship; Redes neurais; Redes convolucionais; Classificação; Detecção; Geração automática; Censura; Pornografia; CIENCIA DA COMPUTACAO::TEORIA DA COMPUTACAO

The adult content available on the internet generates health problems and behavioral disorders. The consumption of pornography is favored by the ease of access, low cost and anonymity of Internet users. Breaking at last one of these factors can minimize the consumption of this content, however, given the volume, it is necessary to analyze the content automatically. In this sense, Deep Learning can perform complex tasks automatically. This thesis attacks the ease of access to pornography by applying automatic censorship through 3 Deep Learning approaches: classification, object detection and automatic generation. In the classification approach, 8 predictive models of different neural network architectures were trained and evaluated, where the predictive results reached accuracy above 99%, processing up to 40 FPS. It was observed that the most significant regions for pornography classification are related to the intimate body parts. The second approach censored pornography with object detection methods. An intimate body parts detection dataset was constructed which allowed the training of models for censoring intimate body parts that achieved mAP = 0.6961. A neural network for detection, called CensorNet, was built, generating promising predictive results. We build CensorPlus, a network composed by a second output for classification. This network creates a hybrid method for object detection and image classification. Finally, the third approach to this thesis presents AttGAN, a method based on image-to-image translation that uses neural networks to generate automatic censorship. The method utilizes attention masks generated by AttNET, a classification-trained neural network converted to generate such masks. Three AttGAN variations were developed, and we designed an online survey where 21 participants compared the results. The results indicate an advantage to the AttGAN+ method, pointed as the best method in 1, 050 opinions collected. The AttGAN+ method was incremented by merging the input image with the censored output, giving rise to the AttGAN++ method, resulting in a censored image that preserves peripheral characteristics of the original image.